Text-to-SQL systems must balance generation accuracy, execution reliability, and inference efficiency while grounding queries in heterogeneous database metadata. We present QueryBridge, an adaptive metadata-aware Text-to-SQL pipeline that combines complexity-aware LLM routing, metadata-grounded generation, business-rule retrieval, selective ReAct-based SQL repair, and pre-execution authorization. On the BIRD development set comprising 1,534 questions across 11 databases, QueryBridge achieves 68.4% execution accuracy, an R-VES of 70.1, and a 97.1% execution success rate, with an average latency of 6.7 s and an estimated inference cost of $0.014 per query. Under the same evaluation setting, it improves execution accuracy over CHESS by 2.3 percentage points while reducing average latency and estimated cost. Component ablations show that metadata grounding provides the largest accuracy gain, followed by ReAct-based repair and business-rule knowledge, while adaptive routing improves the accuracy–efficiency trade-off.

Natural Language Processing,Text-to-SQL,Large Language Models,Data Lake Architecture,Query Processing,AWS Cloud Computing